On Sun, 2013-03-31 at 15:45 -0400, Tom Lane wrote:
> Really, when we're traipsing down a bucket
> list, skipping over bucket entries with the wrong hash code is just
> about free, or at least it's a whole lot cheaper than applying ExecQual.
>
> Perhaps what we should do is charge the hash_qual_cost only for some
> small multiple of the number of tuples that we expect will *pass* the
> hash quals, which is a number we have to compute anyway. The multiple
> would represent the rate of hash-code collisions we expect.
+1.
> I'd still be inclined to charge something per bucket entry, but it
> should be really small, perhaps on the order of 0.01 times
> cpu_operator_cost.
> Or we could just drop that term entirely.
FWIW, either of those are fine with me based on my limited experience.
> Maybe what we should be doing with the bucketsize numbers is estimating
> peak memory consumption to gate whether we'll accept the plan at all,
> rather than adding terms to the cost estimate.
Sounds reasonable.
Ideally, we'd have a way to continue executing even in that case; but
that's a project by itself, and would make it even more difficult to
cost accurately.
Regards,Jeff Davis